PyAF 5.0 Release Process

[antoinecarme]

Time to put in a place a final path/countdown through the next PyAF release, expected 2023-07-14

A branch 5.0-Fixes is already in place and will contain the final touches. The CI process is OK.

https://github.com/antoinecarme/pyaf/tree/5.0-Fixes

Only user comments and fixes are welcome and accepted.

No additional/new feature. Early adopters feedback welcome until July.

Versioning Notes :

1. This version is a major version as in https://semver.org/, like other annual versions (usually released on 20xx-07-14 ;).
2. It introduces a radical change in the model selection processes (Voting, Model Complexity, MAPE -> MASE by default).
3. It may not be backwards compatible in some places (reverting to the old behavior may need tweaking options).

[antoinecarme]

Original planning :

1. Forecasting Models based on Deep Learning Attention Mechanisms => Delayed to PyAF > 7.0
2. Recurrent tasks : More hardware architectures, forecasting performance measures, benchmarks, bug fixes, docs and optimization/profiling. OK.
3. User feedback and issues Ok.

[antoinecarme]

Implemented/Processed So far :

1. Python 3.11 support #227
2. RISC-V Hardware Platform Validation #208
3. Outlier-resistant forecasting Performance Measures #209
4. Use PyTorch as the reference deep learning framework/architecture for future projects #211
5. Investigate Model Esthetics for PyAF #212
6. Investigate Large Horizon Models #213
7. Bad plot for shaded area around prediction intervals in hourly data #216
8. Automate Prototyping Activities - R-based Models #217
9. Failure to build a multiplicative ozone model with Lag1 trend #220
10. Add Differentiable Variant of SMAPE Performance Measure #221
11. Revisit Model Complexity Definition #223
12. Run some Sanity Checks for PyAF 5.0 #224
13. Forecast Quantiles Plots can be improved  #225
14. Re-run the Benchmarking process for PyAF 5.0 #222
15. Use MASE by default for PyAF Model Selection #229 ?
16. PyAF 5.0 Final Touch 1 : discard some non-significant components #230
17. PyAF 5.0 Final Touch 2: Disable alpha in ridge regressions #231
18. Pyaf 5.0 Final Touch 3 : report plot filenames in the logs #232
19. Provide some UML docs for PyAF integrators #233
20. Pyaf 5.0 Final Touch 4 : Add More Tests #234
21. Use MaxAbsScaler for some Multiplicative Signal Transformations #235
22. Pyaf 5.0 Final Touch 5 : Add more info about Exogenous Data Used in ARX Models #236
23. Pyaf 5.0 Final Touch 6 : Disable Timing Loggers by default #237
24. Pyaf 5.0 Final Touch 7 : Improve the Guess of Window Length for Moving Average Trends #238
25. Pyaf 5.0 Final Touch 8 : Use an Optimal Choice Rule for the Quantization Signal transform #239

[antoinecarme]

Created a new release candidate 5.0-rc1.

Will become an official 5.0 on 2023-07-14 if no change until then.

[antoinecarme]

Can be installed with

pip install --upgrade git+https://github.com/antoinecarme/[email protected]

[antoinecarme]

Implemented/Processed So far (for changelog):

1. Python 3.11 Support : Python 3.11 support #227
2. New Hardware Support : RISC-V Hardware Platform Validation #208
3. New Performance Measures : Outlier-resistant forecasting Performance Measures #209, Add Differentiable Variant of SMAPE Performance Measure #221
4. Model Selection Improvement : Investigate Model Esthetics for PyAF #212, Investigate Large Horizon Models #213 , Revisit Model Complexity Definition #223, Use MASE by default for PyAF Model Selection #229
5. Signal Transformation Improvements : Use MaxAbsScaler for some Multiplicative Signal Transformations #235, Pyaf 5.0 Final Touch 8 : Use an Optimal Choice Rule for the Quantization Signal transform #239
6. Generic Modeling : PyAF 5.0 Final Touch 1 : discard some non-significant components #230, PyAF 5.0 Final Touch 2: Disable alpha in ridge regressions #231, Pyaf 5.0 Final Touch 5 : Add more info about Exogenous Data Used in ARX Models #236, Pyaf 5.0 Final Touch 7 : Improve the Guess of Window Length for Moving Average Trends #238
7. Plotting Functions Improvements and Bug Fixes : Bad plot for shaded area around prediction intervals in hourly data #216, Forecast Quantiles Plots can be improved  #225, Pyaf 5.0 Final Touch 3 : report plot filenames in the logs #232
8. New Docs : Provide some UML docs for PyAF integrators #233
9. Bug Fixes : Failure to build a multiplicative ozone model with Lag1 trend #220
10. PyAF "Forecast Tasks" : Use PyTorch as the reference deep learning framework/architecture for future projects #211, Automate Prototyping Activities - R-based Models #217
11. Recurrent Tasks : Re-run the Benchmarking process for PyAF 5.0 #222, Run some Sanity Checks for PyAF 5.0 #224, Pyaf 5.0 Final Touch 4 : Add More Tests #234, Pyaf 5.0 Final Touch 6 : Disable Timing Loggers by default #237

[antoinecarme]

CI tests completed OK as of 2023-04-04 22:26:00

https://app.circleci.com/pipelines/github/antoinecarme/pyaf/394/workflows/b8ee0adb-9690-43d8-90cf-483bd71b3342/jobs/405

[antoinecarme]

Switched to Github Actions CI/CD. Farewell CircleCI !!!

The switch was smooth.

https://github.com/antoinecarme/pyaf/actions/runs/5501701813/jobs/10025475124

[antoinecarme]

DONE.